We're observing an assert in ShenandoahHeapRegion:get_update_watermark():

assert(bottom() <= watermark && watermark <= top(), "within bounds");

The trouble there is that arraycopy routines will get there concurrently with evacuating threads that update update_watermark under lock, but the read-access is not under lock. The observing thread can observe the updates to top and update_watermark in wrong order and thus trip that assert. It happens reliably on aarch64 with the following test:

CONF=fastdebug LANG=C LOG=info make run-test TEST=compiler/types/correctness/OffTest.java TEST_VM_OPTS="-XX:+UnlockExperimentalVMOptions -XX:+UseShenandoahGC -XX:ShenandoahGCHeuristics=aggressive"